Intelligent Grasping with the Robotic Opposable Thumb

Abstract

Recent advances in the field of grasp planning have used heuristics, dimensionality reduction, machine learning, and haptic feedback, with a high degree of success, to plan grasps for simple grippers and/or simple object geometry. We look at applying some of these techniques to the anthropomorphic Meka gripper. First, dimensionality reduction is attempted. We show that dimensionality reduction does not accurately predict the thumb position. A new algorithm is proposed in which measurements from 2D images are used to classify the thumb opposition angle to one of three positions. The remaining joints employ a reactive torque-control strategy to complete the grasp. The algorithm achieves force closure for 82% of 39 household objects. It is simple, computationally fast, and achieves a success rate that is similar to other contemporary grasp planning algorithms.